Recently I wanted to use deep learning to track fridge inventory. The idea was that a camera would track items and then make decisions based on that.
The biggest challenge i faced was that there was no dataset available , and I had to create my own. So I have an idea to try to make computer vision dataset creation less of a pain.
I am thinking of creating a library that allows for the following:
Use a phone camera to create a 3d model of each object you want to create (this technology already exists and is doable)
Project that object in a virtual environment, then take images from different orientation/lighting conditions etc.
My question is, is there a requirement for such a thing ?
I am also looking for people to collaborate with , if I do decide to work on it
I am not looking for a specific dataset to train on. Rather I am asking if there is a requirement for a tool that will CREATE those datasets using the method i mentioned.
I am only aware of Microsoft having proprietary software doing this, would you know and share info about any open source solutions that do this?
What you train your model on introduces “bias” into your model. It would be much easier to generate training datasets this way, but I initially would be suspect it would be too clean.
You may consider hiring people on Amazon’s mechanical turk to take pictures of their fridge with tagged items in the picture and generating a dataset from that. You could open source it to get additional contributors.
What lesson’s throughout part1 and part2 have you found useful to succeed in such a tracking scheme?